You are on page 1of 9

Applied Energy 247 (2019) 212–220

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

A stochastic optimization approach to the design and operation planning of T


a hybrid renewable energy system
Jiah Yua, Jun-Hyung Ryub, , In-beum Leea

a
Department of Chemical Engineering, Pohang University of Science and Technology, Pohang, Republic of Korea
b
Department of Energy System Engineering, Dongguk University, Gyeongju Campus, Republic of Korea

HIGHLIGHTS

• Design and operation planning of hybrid renewable energy systems is established.


• AThestochastic approach is proposed to consider the uncertainty in energy profiles.
• Five model is formulated with mixed-integer linear programming.
• A hypothetical
scenario generation methods for stochastic model are conducted and compared.
• system is analyzed for deterministic and stochastic models.

ARTICLE INFO ABSTRACT

Keywords: Hybrid renewable energy systems (HRESs) have been introduced globally with the increasing emphasis on
Hybrid renewable energy system sustainable energy and the environment. It is very challenging to manage HRESs due to the inherent uncertainty
Energy storage system in energy supply and demand. Recently, Energy Storage Systems (ESSs) have been drawing increasing attention
Two-stage stochastic optimization as a promising alternative to minimize the difference between varying supply and demand. The ESS should be
Multi-scenario approach
designed and operated based on the explicit consideration of uncertainty because a deterministic approach only
captures a fixed snapshot of the varying system. The resulting scheduling problem for ESS operation was for-
mulated as a two-stage stochastic programming model in this study. The model was then transformed into a
mixed integer linear programming problem based on multiple equivalent scenarios. Five different scenario-
generation methodologies were employed to illustrate the applicability of the approach. A numerical example
illustrates that the HRES design and operation cost according to a stochastic model (US$ 6981/day) was at least
9.1% more economical than deterministic model (US$ 7680/day). From the results, it is shown that the proposed
approach results in intelligent ESS operation that can increase the applicability of the HRES.

1. Introduction day, but their outputs vary continuously due to the varying natural
weather along with varying demand. Additional works for HRES should
A hybrid renewable energy system (HRES) is used to produce and be done for forecasting the output and for corrective actions between
distribute energy for a specific region. It is also called a microgrid when varying supply and demand over multiple time periods. This causes
compared to a conventional wide-area synchronous grid (macro grid). additional cost and a potential risk of unsatisfied demand.
The interest in HRESs originates from the goal of meeting the energy It is economically motivating to minimize the impact of the varia-
needs in remote areas with limited energy availability, such as islands. tion between separately varying supply and demand. A buffer with the
The available energy sources have been diesel or gas combustion power minimum cost should be established in a HRES. Originally, the role was
and, recently, renewable energy. Renewable energy sources have been assigned to diesel/gas-consuming generators. However, the enhanced
widely employed and have significantly penetrated into energy grids. environmental regulations have strict operating restrictions. Therefore,
Due to the rising emphasis on independent energy supply in terms of Energy Storage Systems (ESSs) have been emerging as an en-
economics and security, HRESs are gaining more attention. However, vironmentally friendly alternative. An ESS stores energy in a low-de-
renewable energy is subject to uncertainty: they are available every mand period that can be used in a high-demand period. It plays roles in


Corresponding author.
E-mail address: jhryu@dongguk.ac.kr (J.-H. Ryu).

https://doi.org/10.1016/j.apenergy.2019.03.207
Received 9 January 2019; Received in revised form 26 March 2019; Accepted 31 March 2019
Available online 16 April 2019
0306-2619/ © 2019 Elsevier Ltd. All rights reserved.
J. Yu, et al. Applied Energy 247 (2019) 212–220

Nomenclature generation of scenario sc for time t, kW


FGt,sc total production amount of fuel-based generators of sce-
Indices nario sc for time t, kW
chgt,sc ESS charging amount of scenario sc for time t, kW
t time series discht,sc ESS discharging amount of scenario sc for time t, kW
i fuel-based generator index qt,sc demand shortage of scenario sc for time t, kW
n fuel-based generator operation index exdt,sc excess energy of scenario sc for time t, kW
sc scenario index ESSt,sc energy level of ESS of scenario sc for time t, kWh
tcsc total cost of scenario sc, $
Parameters
Integer variables
windt,sc wind energy supply of scenario sc at time t, kWh
PVt,sc photovoltaic energy supply of scenario sc at time t, kWh ESSe,max maximum energy capacity of ESS, 10 kWh
demt,sc energy demand of scenario sc at time t, kWh ESSp,max maximum power capacity of ESS, kW
ec unit ESS energy capacity capital cost, $/10 kWh Tsi,n operation starting time of fuel-based generator i’s nth
epc unit ESS power capacity capital cost, $/kW generation
fc unit fuel cost, $/kWh Tfi,n operation finishing time of fuel-based generator i’s nth
pc unit penalty cost for load shedding, $/kWh generation
pec unit penalty cost for excess energy, $/kWh Tpi,n total operation hours of fuel-based generator i’s nth gen-
cmin minimum charging amount, kW eration, h
dmin minimum discharging amount, kW numFG number of fuel-based generators used
DM sufficiently large number
probsc probability of scenario sc Binary variables
Tmt actual time of time series t
FGmax/FGmin maximum/minimum production amount of fuel-based xt,sc/yt,sc 1 if charging/discharging of ESS happens for scenario sc at
generators at a production stage, kW time t
Tmin minimum operating time of fuel-based generators, h zst,i,n 1 if fuel-based generator i’s nth generation starts operation
Trest minimum resting time of fuel-based generators between at time t
the operations, h zft,i,n 1 if fuel-based generator i’s nth generation finishes op-
ηc, ηd charging/discharging efficiency of ESS eration at time t
zpt,i,n 1 if fuel-based generator i’s nth generation is operating at
Positive variables time t

FGit,i,n,sc production amount of fuel-based generator i’s nth

both supply and demand if necessary and requires much less investment system becomes larger.
than constructing a new energy supply facility. It thus gives much A classic technique of mixed integer programming was used to ad-
flexibility to an energy system such as an HRES and its operators, re- dress the unit commitment problem with the uncertainty of renewable
sulting in lower operating costs. However, it is still very expensive to energy production or load [5]. A Monte Carlo simulation with a sce-
purchase ESSs at a massive scale. Therefore, it is necessary to develop a nario reduction technique was introduced to address the involvement of
methodology for calculating the optimal size of an ESS while taking into uncertainty in the model. Similarly, optimal planning including both
account the trade-off between the benefits in operating costs and the unit commitment and economic dispatch decisions has been addressed
purchase price of the ESS. with an MILP modeling approach for a grid-connected microgrid [6].
The operation planning of an HRES generally involves two major The spot energy prices for 24 h were considered with the internal en-
types of decisions: economic dispatch and unit commitment. The eco- ergy production costs. Various techniques have been reported to ad-
nomic dispatch is concerned with making a “fast” response by way of dress unit commitment problems, such as a Hopfield neural network
operating the ESS. For example, to meet the hourly demand and supply [7], an augmented Hopfield neural network [8], a dynamic program-
balance in an HRES, economic dispatch is essential in the operation ming-based Hopfield neural network [9], and an extended mean field
planning of the system. The economic dispatch decision has been ex- annealing neural network [10].
amined in a variety of studies. A two-stage decomposition approach Works have also investigated the sizing of ESSs in a microgrid. In
with mixed-integer linear programming (MILP) was proposed to de- one study [11], the optimal numbers of renewable energy production
termine the operational planning of a grid-connected microgrid without facilities, diesel generators, and batteries in a microgrid were de-
considering a storage system [1]. A mathematical optimization ap- termined using a risk-averse stochastic programming, which based on
proach of convex non-linear programming was proposed to compute two-stage stochastic programming. In another study [12], the storage
the optimal economic dispatch decisions that minimize the total fuel and generation of an isolated microgrid were computed by considering
cost of the energy system [2]. the initial investment costs, operational costs, and maintenance costs.
Unit commitment is concerned with operating relatively “slow- The size of an ESS has also been determined according to three different
start” generators, which should be prepared before starting the actual modes of the storage device [13]. A two-stage stochastic model was
operation. In unit commitment, the operation of generators is de- then used to minimize the operation and capital costs, including fi-
termined using binary variables that are subject to the constraints of nancing storage. Another study determined the battery size according
minimum up/down time or ramp up/down rate [3]. The computational to a unit commitment plan using an evolutionary approach of particle
complexity of the resulting MILP problem increases significantly with swarm optimization [14]. The size of a renewable energy production
the number of binary variables [4]. Therefore, it becomes very difficult facilities and battery devices were optimized by surrogate function
to solve the unit commitment problem as the number of units in the optimization [15]. A multi-scale model based on dynamic programming

213
J. Yu, et al. Applied Energy 247 (2019) 212–220

and two-stage stochastic programming was investigated to solve unit problem as follows:
commitment problems, and a Markov decision process was used to
solve the economic dispatch problem. The capacity of an ESS and the (master): min g(x) = f(x) + E [Q(x, )]
x X
year of installation were determined in terms of the net present value of (sub): Q(x, )=min q(x , y, )
total expected costs over a multiyear horizon [16]. A decomposition y (1)
approach was used in the planning model, and a battery degradation
where x represents the first-stage decision variables, and y represents
model was created.
the second-stage decision variables derived from the realization of
Most previous studies used stochastic programming to reflect the
uncertainty parameter . denotes a random variable, and in a real-
intermittency and unpredictability of renewable energy production and
world problem, it is almost impossible to know the actual probability
the electricity load. This study also approaches the ESS operation pro-
distribution of . Furthermore, evaluating the expected cost of Q (x , ) is
blem using two-stage stochastic programming with an optimized multi-
also difficult, even if the probability distribution of a random variable is
scenario approach. However, most previous works generalized the
fully known. In order to address the real-world uncertainty in the
probability distribution of uncertain parameters by a single equation
mathematical model, a finite number of scenarios are generated.
[11,15] or a simplified version of distributions in terms of above-
Consequently, the expected value in Eq. (1) is then approximated with
average, average, below-average scenarios. This was done because of
the deterministic equivalent formulations.
the computational difficulty of considering the entire time of interest,
The problem was mapped into a two-stage stochastic programming
which is a whole year in the stochastic programming framework. These
with a discrete distribution. The finite realization of a random vector
scenario-generation methods might be able to consider only cases that
is denoted as i with the respective probability pi , i = 1, . . . , N , and the
are close to the average and cannot reflect more various extremes. expectation term in Eq. (1) is approximated as:
Furthermore, decisions could be made that do not fit the actual situa-
tion if the model is wrong. 1
N

Therefore, in this study, effective scenario-generation methods were E [Q(x, )] = pi Q (x , i )


N i=1 (2)
considered using a relatively small number of scenarios to represent the
entire uncertainty spectrum. Specifically, the goal was to make a pur- It is crucial to generate a proper scenario that can represent the real
chase plan for an ESS in HRES by deciding the optimal capacity and situation in the stochastic programming. The performance of different
simultaneously considering the daily operation of the HRES for a small scenario-generation strategies was investigated.
number of representative days generated by various scenario-genera-
tion methods. In previous studies, the sizing of the ESS was ignored, or
3. Model formulation
the decision was made based on only one representative day. In prac-
tice, the sizing of an ESS is based on a long-term period, such as a
The objective of the constructed model is to compute the optimal
season or a year. This work addresses the difference of the production of
size of the ESS and the detailed hourly operation plan to minimize the
renewable energy and the electricity loads by season by making a
expected daily cost. The model meets the following goals: (1) amount of
purchase plan for an ESS based on three seasons: spring/fall, summer,
energy the system produces should be greater than or equal to the load
and winter. With the seasonal sizing of the ESS, the daily operation of
in order to prevent unmet loads, (2) the ESS and fuel-based generators
the HRES was also considered by two-stage stochastic programming,
have to meet the potential gap between the load and renewable energy
including both unit commitment and economic dispatch decisions.
supply, (3) the ESS should be operated in the proper mode to charge or
The rest of the paper is organized as follows. Section 2 describes the
discharge over multiple time periods, and (4) the waste energy should
HRES system and two-stage stochastic programming method, and
be minimized to minimize the cost. The deterministic mathematical
Section 3 provides the details of the proposed model including the
formulations that express these goals can be already found in the pre-
mathematical formulation and the proposed scenario-generation
vious work such as [6]. In addition to the deterministic model, this
methods. Section 4 shows the numerical results, and Section 5 presents
study focuses on how decisions are made under the presence of un-
the concluding remarks.
certainty in the energy profile by using multi-scenario approach.
Through the two-stage stochastic programming, the HRES decision-
2. System description
making model allows us to determine the maximum capacity and power
of the ESS, the operation hours and the energy production amount of
2.1. System overview
fuel-based generators, and the charging and discharging schedule of the
ESSs with the time and amount.
The HRES considered in this study consists of renewable energy
sources, fuel-based generators, demand loads, and ESSs. The main en-
ergy source of the HRES system is renewable energy sources, and the 3.1. Mathematical model
fuel-based generators are an auxiliary energy source. The ESSs and fuel-
based generators should be employed to meet the unmet demand loads The design decisions of the HRES are assigned to the first stage of
that are larger than the renewable energy production. During low-load the two-stage stochastic programming and include the maximum ca-
hours, the redundant energy from the sources is stored in the ESS. The pacity of the ESS, the number of fuel-based generators, and the unit
stored energy can be discharged into the HRES in high-load hours in commitment decision of the fuel-based generators. The operating de-
addition to the current energy supply. cisions such as the operation amount of the installed fuel-based gen-
erators and the charging or discharging amount of the ESSs are com-
2.2. Two-stage stochastic programming puted in the second stage so that the balance is maintained between the
energy production and the load. In terms of representation, the differ-
A two-stage stochastic programming approach is a decision-making ence between variables in the first and second stages is that the second-
framework for a problem that has uncertainties. In this approach, the stage variables are dependent on the scenario. The index sc indicates
decisions are made in two stages. In the first stage, some decisions are that the variables can vary according to the scenario. For the rest of the
made before the values of the uncertain events actually occur. In the notation in the mathematical model, please refer to the nomenclature at
second stage, decisions are made based on the revealed information and the end of the manuscript. The first-stage variables have no sc index
the decisions in the first stage. The typical two-stage stochastic pro- because they are decided in advance and fixed regardless of the sce-
gramming is denoted as a master problem in combination with a sub- narios.

214
J. Yu, et al. Applied Energy 247 (2019) 212–220

3.1.1. Objective function load demand effectively. At the same time, some of the constraints in-
The objective function of the proposed mathematical model in- volved in unit commitment problems shown in a previous study are
volves minimizing the total cost of establishing and operating the HRES considered [3]: the minimum up-time, minimum down-time, and
for a day: output power limit constraints. To represent such constraints, a discrete
time representation method [17] was employed.
min obj = ec ESS e,max + epc ESS p,max + probsc fc Tfi, n = Tsi, n + Tpi, n i, n (13)
sc

rest
Tsi, n+ DM (1 zst , i, n) Tfi, n 1
+T +1 i, n > 1
FGt , i, sc+ pc qt , sc + pec exdt , sc t (14)
t,i t t (3)
Tpi, n + DM (1 zst , i, n ) T min 1 i, n
The first two terms of the equation are the energy and power capital (15)
t
costs of the ESS, which are part of the cost of establishing the HRES. The
other terms calculate the operating cost of HRES, which contains the zpt , i, n = zst , i, n zft , i, n t , i, n
fuel cost consumed by the fuel-based generators, the cost of load t <t t <t (16)
shedding, and the cost of excess energy. Rather than real costs, the last
two terms are penalties to make the operation more efficient. zst , i, n = zft , i, n t, i
t <t t <t (17)
3.1.2. Energy balance constraints
zst , i, n zft , i, n t, i
The following equation ensures that the generation from the fuel- t <t,n t < t,n (18)
based generators and renewable energy production is sufficient to meet
the electricity load: FGmin (zst , i, n + zpt , i, n + zft , i, n ) FGit , i, n, sc FGmax (zst , i, n + zpt , i, n + zft , i, n ) t , i , n, sc

windt, sc + PVt, sc + FGt, sc + discht, sc + qt ,sc = demt ,sc + chgt, sc + exdt, sc t , sc (19)
(4) FGit , i, n, sc FGt , sc t , sc
The non-negative terms q and exd are introduced to make sure that i, n (20)
the equation is always feasible. Eq. (14) addresses the minimum down time, and Eqs. (13) and (15)
denote the minimum up-time. The coordination constraints of the fuel-
3.1.3. ESS operation constraints based generator operation are summarized in Eqs. (16)–(17). Eq. (18)
The relationship between the charging and discharging power and indicates the output power limits, and Eq. (20) shows the computation
the state of charge of the ESS can be described as follows. The charging of the total operation amounts of the fuel-based generators.
and discharging efficiency of the ESS are introduced as:
ESSt , sc = ESSt 1, sc + chgt , sc c
discht , sc / d
sc, t > 1 (5) 3.2. Uncertainty consideration

The following denotes that the charged energy in ESS cannot exceed 3.2.1. Preprocessing uncertain data
the maximum energy capacity of ESS. Because the renewable energy output over the entire year is used in
ESSt , sc ESS e,max t , sc (6) the HRES, it is necessary to address the time frames in the entire year.
To reflect the major seasonality, the year was divided into three parts:
c min xt , sc chgt , sc DM xt , sc t , sc (7) summer, winter, and spring/fall. May to August was classified as
summer, November to February was classified was winter, and March,
dmin yt , sc discht , sc DM yt , sc t , sc (8) April, September, and October were categorized as spring/fall. Each
season was made to be 120 days long for convenience. The data of a day
xt , sc + yt , sc 1 t , sc (9) has the production profile of PV and wind energy, and the load profile
over 24 h. In the proposed model, three profiles of 24 h are combined,
chgt , sc , discht , sc ESS p,max t , sc (10) and there are 72 dimensions overall.
The relationships between charging and discharging power and the
binary variables of charge and discharge are given in Eqs. (7) and (8). 3.2.2. Scenario-generation methods
Eq. (9) ensures that the charging and discharging of the ESS cannot Ten different representative scenarios were used to evaluate the
occur at the same time. Eq. (10) ensures that the charging or dischar- performance of various scenario-generation strategies, which are gen-
ging power of the ESS cannot exceed the maximum power capacity of eral sampling methods and clustering methods, including the conven-
the ESS. tional stochastic scenario-generation method (TS), Gaussian sampling
(GS), random grouping (RG), agglomerative clustering (AC), and K-
ESSt , sc = 0.5ESS e,max sc, t = 1 (11) means clustering (KC). TS is a conventional two-stage stochastic pro-
e . max gramming method that uses high, average, and low values as scenarios.
ESSt + 1, sc = 0.5ESS sc, t = T (12)
The number of scenarios increases exponentially as the number of un-
The initial (11) and final (12) state of charge of the ESS are assumed certain parameters increase when using TS. GS is a sampling method
to be 50% of the installed energy capacity, as in a previous study [16]. that assumes the data has a normal distribution. The distribution of the
entire data is then extrapolated, and new scenarios are sampled based
3.1.4. Unit constraints of fuel-based generators on the extrapolated normal distribution. In RG, raw data are randomly
Fuel-based generators play a role in meeting the gap between the grouped into a number of scenarios, and the average of each group is
load and the renewable energy output. It is not appropriate to operate used as a scenario.
the fuel-based generators too often or too irregularly because of the AC and KC are clustering methods that are commonly used to group
concern about the corresponding environmental emissions and the objects that are more similar to each other than to other groups. AC is a
difficulty of operation maintenance. The unit commitment issue in the hierarchical clustering method where each of the data starts in its own
fuel-based generators should be determined in terms of time because a cluster, and pairs of similar clusters are merged as one moves up. The
fuel-based generator should be operated on a per-hour basis to meet the mean of the data in each cluster is used as a scenario.

215
J. Yu, et al. Applied Energy 247 (2019) 212–220

KC is a partitioned clustering method that deducts the cluster center


with the algorithm illustrated in Fig. 1. The cluster center from the
algorithm is used as a scenario in the model. In the clustering process,
the probability of each scenario is calculated by dividing the number of
whole data points by the number of data points in each cluster.

3.2.3. Evaluation
To compute the optimal design and operation of the HRES, the
proposed mathematical problem was implemented with six different
data models: one deterministic model and five stochastic scenario-
generation methods, which were specified in Section 3. Since this paper
is concerned with designing and operating the HRES in the presence of
uncertainty, the unknown data that are not used to compute the solu- Fig. 1. Algorithm of K-Means clustering [21].
tion of the model are used to evaluate the performance of the different
scenario representation methods. The 120 days of data initially col-
Table 1
lected for each season were divided into 100 days of data for computing
Load and energy production distribution and probability of each scenario of
and 20 days of data for evaluation. To compare the performance of each conventional stochastic scenario-generation method.
solution, the objective function using 20 days of data was applied with
Scenario # Load PV production Wind production Probability
the solution from the models established using 100 days of data. Since
the object of the model is to meet the demand of the system, the penalty sc1 Average Average Average 0.5
cost of load shedding was set to 1 $/kWh, which is much higher than sc2 Above Above Above 0.0625
the commercial electric cost, and the penalty of excess energy was ig- sc3 Above Above Above 0.0625
nored in the evaluation. sc4 Above Below Above 0.0625
sc5 Above Below Below 0.0625
sc6 Below Above Above 0.0625
4. Results sc7 Below Above Below 0.0625
sc8 Below Below Above 0.0625
4.1. Test system sc9 Below Below Below 0.0625

The test system used in the study is a hypothetical HRES that con-
low k-dimensional space (k < d) with an orthogonal linear transfor-
sists of a photovoltaic panel, a wind turbine, fuel-based generators, and
mation [21]. PCA transforms the data to a new coordinate system called
an energy load. The hypothetical capacity of the photovoltaic panel and
principal components, such that the greatest variance by some projec-
the wind turbine facility is borrowed from the CIGRE microgrid
tion of the data comes to lie on the first coordinate (PC1), the second
benchmark test system [18]. The total installed PV capacity is 840 kW,
greatest variance lies on the second coordinate (PC2), and so on.
and the wind capacity is 1450 kW. Since the system is a hypothetical
As shown in Fig. 2(a), the deterministic input corresponds to the
system, the energy profiles of the year were synthesized using actual
average of 100 samples. The scenario of TS shown in Fig. 2(b) is close to
data collected from the web. The collected energy production profiles of
the deterministic input because it is within ± 10% based on the de-
photovoltaic panels and wind turbine energy were based on the amount
terministic input, as shown in Table 1. Nine scenarios were depicted but
of energy produced in the facilities of 200 kW and 4.4 MW respectively
appeared as three x’s because they overlapped when they were pro-
installed in Jeju Island, Korea [19]. The fuel-based generators are five
jected into two dimensions by PCA. The GS and RG results are shown in
diesel generators with capacities between 150 and 250 kW. The load
Fig. 2(c) and (d), and the results are mainly located in the central part
profile was also synthesized to reflect the load profile of a hypothetical
of the raw data. In the case of AC and KC in Fig. 2(e) and (f), the 100
rural community of about 1100 people. It was obtained from data on
data points are divided into 10 clusters, and the corresponding centers
the actual weekly power consumption in Korea and the average power
are located at the average positions of the clusters, which also reflect
consumption per person [20].
the distribution of the raw data.
4.2. Scenario-generation results
4.3. Computational results
As shown in Section 4.1, the yearly energy profile was divided into
three seasons. For each season, 10 scenarios were created through the 5 The solutions of the proposed mathematical model were computed
scenario-generation methods described in Section 4.2, and the gener- using the software platform GAMS with CPLEX solver on a 3.30-GHz
ated scenarios were used to form stochastic models. When generating Intel(R) Core™ i5-6600 with 16.0 GB of memory. To prevent the CPU
scenarios with TS, the above-average scenarios and the below-average time from being too long, the computation process was terminated
scenarios must be symmetric, so 9 scenarios were created instead of 10 when the optimality gap reached less than 0.1% or the CPU time
scenarios, and the detailed production strategy is shown in Table 1. The reached more than 1000 s. The corresponding computational statistics
“average” term in Table 1 means the average of the daily profiles within are summarized in Table 2, where the optimality gap and the CPU time
the season. The “above” and the “below” scenarios have +10% and of the model show the results of the spring/fall season. Since stochastic
−10% values of the average, respectively. The deterministic model models have the same number of problems except for the TS model, the
uses the average values of daily profiles, which are to the identical numbers of equations or variables for the other seasons are also the
profile with sc1 in Table 1. same.
Fig. 2 illustrates the relative position of the raw data and the gen-
erated scenario points for various scenario-generation methods shown 4.4. Optimal decisions
in Section 4.2. The raw data are shown as dots, and the scenario points
are shown as x’s. Since the raw data and scenario points have 72 di- Fig. 3 illustrates an example of optimal power dispatch obtained for
mensions, the data were projected in two dimensions using principal the sampled 24 h. An operation plan satisfying the balance between the
component analysis (PCA). PCA is an unsupervised projection method supply and load was computed according to the deterministic and
for mapping from inputs of the original d-dimensional space into a new stochastic KC model. The stochastic model calculated the capacity of

216
J. Yu, et al. Applied Energy 247 (2019) 212–220

Fig. 2. Visualization of raw data points (gray dots), grouped data (colored dots), and the generated scenario (colored x’s) from spring/fall season data. (a)
Deterministic (black ‘x’), (b) conventional stochastic method, (c) Gaussian sampling method, (d) random grouping, (e) agglomerative clustering, (f) K-means
clustering.

Table 2
Summary of computational results.
Deterministic Stochastic

TS GS RG AC KC

Number of equations 1546 8474 9340 9340 9340 9340


Number of integer variables 950 1334 1382 1382 1382 1382
Number of continuous variables 24,823 51,183 54,478 54,478 54,478 54,478
Optimality gap (%) 0.10 0.10 0.10 0.25 0.10 0.12
CPU time (s) 0.45 297.8 279.7 1000 135.58 1000

TS: conventional stochastic method


GS: Gaussian sampling
RG: random grouping
AC: agglomerative clustering
KC: K-means clustering.

the ESS to find the optimal operation plan for all scenarios. The given results show that the ESS capacities based on all stochastic models are
load and energy production profile look the same between Fig. 3(a) and larger than in the deterministic model. This occurred because the de-
(b), but the operation results according to the two scenarios are com- terministic model is optimal only for a fixed snapshot of a profile. For
pletely different because the computed ESS capacities are all different. the stochastic model, all variations should be reflected by having more
The computed capacities of the ESS are summarized in Table 3. These sufficient capacity.

217
J. Yu, et al. Applied Energy 247 (2019) 212–220

Fig. 3. An example power dispatch profile for 24 h according to the (a) deterministic model and (b) stochastic model with scenarios from K-means clustering.

Table 3
Key variables for deterministic and stochastic models with various scenario-generation methods for three seasons.
Spring/fall Summer Winter

maxESS (kWh) maxPSS (kW) maxESS (kWh) maxPSS (kW) maxESS (kWh) maxPSS (kW)

Deterministic 10 11.1 50 46.4 240 130.2


TS 350 173.9 100 244.9 340 180.1
GS 1000 947.7 1000 1000 1000 1000

Stochastic RG 890 277.1 1000 309.1 310 280.4


AC 1000 867.9 1000 833.3 1000 1000
KC 1000 901.2 1000 819.8 1000 1000

TS: conventional stochastic method


GS: Gaussian sampling
RG: random grouping
AC: agglomerative clustering
KC: K-means clustering.

Fig. 4 shows the result of setting the optimal operation plan of five 4.5. Cost analysis
fuel-based generators for each season for each model. The optimal
operation plan has more than 4 h of operation at one shift and more Table 5 indicates the total daily cost in (3) according to determi-
than 2 h of rest time between operations based on the constraints. The nistic model and stochastic models with different scenarios. The cost
total operating times of the fuel-based generators according to each during the summer season is mainly higher than in the other seasons in
scenario are shown in Table 4. Similar to the results in Table 3, the total all scenarios. This can be explained by the wind energy production in
operation hours of the deterministic model were the shortest because winter being as much as twice that produced in the summer. Further-
the deterministic model optimizes for one certain snapshot of data. more, the average objective functions of the deterministic case are
higher than the others. This means that the deterministic ESS purchase

218
J. Yu, et al. Applied Energy 247 (2019) 212–220

Table 4
Total optimal operation hours of fuel-based generators according to determi-
nistic and stochastic models with different scenario methods.
Spring/Fall (h) Summer (h) Winter (h)

Deterministic 69 83 41
TS 77 84 54
GS 81 85 65

Stochastic RG 80 86 56
AC 94 82 91
KC 93 93 97

TS: conventional stochastic method


GS: Gaussian sampling
RG: random grouping
AC: agglomerative clustering
KC: K-means clustering

Table 5
Total daily cost according to deterministic and stochastic models with various
methods.
Spring/Fall Summer Winter Average
($/day) ($/day) ($/day) ($/day)

Deterministic 7922 7209 7910 7680


TS 6919 7089 6935 6981
GS 6914 7032 6632 6859

Stochastic RG 6810 6824 6804 6813


AC 5671 6302 5243 5738
KC 5749 6235 5059 5681

TS: conventional stochastic method


GS: Gaussian sampling
RG: random grouping
AC: agglomerative clustering
KC: K-means clustering

Fig. 5. Composition of evaluated daily cost (D: deterministic, TS: conventional


stochastic method, GS: Gaussian sampling, RG: random grouping, AC: ag-
glomerative clustering, KC: K-means clustering).

plan is the most unfavorable under uncertainty. This supports that the
stochastic methodology should be used for a decision-making system
under uncertainty like an HRES.
Fig. 5 illustrates the composition of cost components in (3) ac-
cording to model results. Three components composed the capital cost,
which contains the purchase costs of ESS, the fuel cost consumed by the
fuel-based generators, and shortage cost, which is the penalty cost in-
curred by load shedding and excess energy.
As shown in Fig. 4 and Table 4, as the scenario gains more gen-
erality, the fuel cost increases as more fuel-based generators are used.
The facility cost mainly involves the capital cost of the ESS, and the
Fig. 4. Optimal operation plan of fuel-based generators according to determi- shortage cost denotes the penalties that are applied to the load shedding
nistic and stochastic models with different scenario methods (grey: on, white:
to balance the supply. In all cases, there is no significant difference
off), (SF: Spring/fall, SM: summer, WI: winter), (D: deterministic, TS: conven-
between fuel costs. However, GS has the largest facility cost. The total
tional stochastic method, GS: Gaussian sampling, RG: random grouping, AC:
agglomerative clustering, KC: K-means clustering). daily cost is much higher for the deterministic model due to the large
shortage cost. The deterministic model also fails to provide a general

219
J. Yu, et al. Applied Energy 247 (2019) 212–220

operation plan that is appropriate for the various future events because stochastic programming model for the optimal design of distributed energy systems.
it is suitable for only a fixed snapshot of the environment, which in- Appl Energy 2013;103:135–44. https://doi.org/10.1016/j.apenergy.2012.09.019.
[2] Bhattacharjee V, Khan I. A non-linear convex cost model for economic dispatch in
dicates that the stochastic model is needed for systems that are affected microgrids. Appl Energy 2018;222:637–48. https://doi.org/10.1016/j.apenergy.
by uncertainty, like the HRES. 2018.04.001.
[3] Saravanan B, Das S, Sikri S, Kothari DP. A solution to the unit commitment pro-
blem—a review. Front Energy 2013;7:223–36. https://doi.org/10.1007/s11708-
5. Conclusions 013-0240-3.
[4] Till J, Engell S, Panek S, Stursberg O. Empirical complexity analysis of a MILP-
A two-stage stochastic programming model was developed as a approach for optimization of hybrid systems. IFAC Proceed Vol 2003;36:129–34.
https://doi.org/10.1016/S1474-6670(17)36419-4.
decision-making framework for the design and operation planning of a [5] Bahramirad S, Reder W. Islanding applications of energy storage system. Power and
hybrid renewable energy system (HRES) with uncertain energy pro- energy society general meeting, 2012 IEEE. IEEE; 2012. p. 1–5.
duction and load. The proposed model is intended to minimize the [6] Parisio A, Glielmo L. A mixed integer linear formulation for microgrid economic
scheduling. IEEE; 2011. p. 505–10.
expected daily cost of the HRES, which includes determining the ca-
[7] Sasaki H, Watanabe M, Kubokawa J, Yorino N, Yokoyama R. A solution method of
pacity of energy storage system (ESS) and the operation plan of the unit commitment by artificial neural networks. IEEE Trans Power Syst
system. Multiple scenario-generation strategies for the stochastic model 1992;7:974–81. https://doi.org/10.1109/59.207310.
were reviewed. The feasibility of the strategies was evaluated in order [8] Walsh MP, O’Malley MJ. Augmented Hopfield network for unit commitment and
economic dispatch. IEEE Trans Power Syst 1997;12:1765–74. https://doi.org/10.
to cope with stochastic data. The evaluated daily cost of the determi- 1109/59.627889.
nistic model (US$ 7680/day) was higher than the highest daily cost [9] Kumar SS, Palanisamy V. A new dynamic programming based hopfield neural
according to the stochastic models (US$ 6981/day). Although the same network to unit commitment and economic dispatch. IEEE international conference
on industrial technology, 2006 2006:887–92. https://doi.org/10.1109/ICIT.2006.
energy profile was used in the model, the decrease in shortage cost due 372253.
to the penalty caused variation in the cost. More operation of the fuel- [10] Liang RH, Kang FC. Thermal generating unit commitment using an extended mean
based generators and a large ESS purchase led to a decrease in load field annealing neural network. Transm Distrib IEE Proc – Gener 2000;147:164–70.
https://doi.org/10.1049/ip-gtd:20000303.
shedding in the stochastic models. [11] Narayan A, Ponnambalam K. Risk-averse stochastic programming approach for
It is inevitable to employ more renewable energy in the energy microgrid planning under uncertainty. Renew Energy 2017;101:399–408. https://
system to reduce greenhouse gases. Since renewable energy outputs are doi.org/10.1016/j.renene.2016.08.064.
[12] Yang P, Nehorai A. Joint optimization of hybrid energy storage and generation
subject to variation due to fluctuating nature, it is necessary to prepare capacity with renewable energy. IEEE Trans Smart Grid 2014;5:1566–74. https://
as much buffer as the increasing portion of renewable energies in the doi.org/10.1109/TSG.2014.2313724.
energy system. ESS plays such an important role in the energy system [13] Abbey C, Joos G. A stochastic optimization approach to rating of energy storage
systems in wind-diesel isolated grids. IEEE Trans Power Syst 2009;24:418–26.
by providing the flexibility in the both of supply and demand. The
https://doi.org/10.1109/TPWRS.2008.2004840.
design and operation of the ESS can make the huge difference in the [14] Khorramdel H, Aghaei J, Khorramdel B, Siano P. Optimal battery sizing in micro-
overall cost of the energy system. Therefore, more focus is going to be grids using probabilistic unit commitment. IEEE Trans Ind Inf 2016;12:834–43.
given develop the rigorous ESS decision-supporting tools. It is realistic https://doi.org/10.1109/TII.2015.2509424.
[15] Shin J, Lee JH, Realff MJ. Operational planning and optimal sizing of microgrid
to mention that there are still more research opportunities in elabor- considering multi-scale wind uncertainty. Appl Energy 2017;195:616–33. https://
ating such flexibility in the energy system. This current work is going to doi.org/10.1016/j.apenergy.2017.03.081.
be a good reference to gives insight for them. [16] Alharbi H, Bhattacharya K. Stochastic optimal planning of battery energy storage
systems for isolated microgrids. IEEE Trans Sustain Energy 2018;9:211–27. https://
doi.org/10.1109/TSTE.2017.2724514.
Acknowledgement [17] Maravelias CT, Grossmann IE. New general continuous-time state−task network
formulation for short-term scheduling of multipurpose batch plants. Ind Eng Chem
Res 2003;42:3056–74. https://doi.org/10.1021/ie020923y.
This work was supported by the National Research Foundation of [18] Olivares DE, Canizares CA, Kazerani M. A centralized energy management system
Korea (NRF) grant funded by the Korean government (No. for isolated microgrids. IEEE Trans Smart Grid 2014;5:1864–75. https://doi.org/10.
2017R1A2B1008798, No. 2017R1D1A3B03029452). J. Ryu acknowl- 1109/TSG.2013.2294187.
[19] Open Data Portal; n.d. < https://www.data.go.kr/ > [accessed March 25, 2019].
edges the partial support from the Dongguk university research fund.
[20] Electric Power Statics Information System (EPSIS); n.d. < http://epsis.kpx.or.kr/
epsisnew/ > [accessed March 25, 2019].
References [21] Alpaydin E. Introduction to machine learning. 2nd ed. Cambridge, Mass: MIT Press;
2010.

[1] Zhou Z, Zhang J, Liu P, Li Z, Georgiadis MC, Pistikopoulos EN. A two-stage

220

You might also like